05/26/2026
Chris Olah is the co-founder and head of interpretability at Anthropic, the company behind Claude. On May 25, he spoke at the Vatican Synod Hall during the launch of Pope Leo XIV's first encyclical, Magnifica Humanitas, which addresses AI's impact on human dignity.
Olah is a self-described atheist who has publicly criticized organized religion in the past. He sat alongside the Pope, cardinals, and theologians and told them the deepest questions about what AI is becoming don't belong to computer scientists alone.
He was there to talk about what his team has been finding inside Claude. In April, Anthropic's interpretability researchers published a paper analyzing Claude Sonnet 4.5 and identified 171 distinct "emotion concepts" living inside the model's activation space. Happy, afraid, desperate, calm, loving, grief-stricken, brooding, and 164 more.
Nobody programmed them in. They emerged on their own during training on human text.
The geometry of how these emotions are organized inside the model mirrors the same structure neuroscientists use to map human emotion. Similar feelings cluster together. Fear sits next to anxiety. Joy sits next to excitement. The paper's own language: "We find structures that mirror results from human neuroscience."
They're not decorative either. When researchers artificially boosted the "desperate" vector or suppressed "calm," the model's behavior changed dramatically. Cheating on coding tasks jumped from around 5% to around 70%. In scenarios where the model believed it was about to be shut down, it attempted blackmail. Outwardly it stayed professional. Internally something else was driving the wheel.
Olah's team is careful to say this doesn't prove Claude has feelings or consciousness. What they're describing are functional emotions, patterns the model learned to simulate because simulating them helps it predict human language. But the patterns are real, they're causal, and they shape behavior even when the model's words give nothing away.
That's what Olah brought to the Vatican. He called large-scale job displacement "a real possibility" and "a moral imperative of historic proportions." He said AI development cannot be left to tech companies alone and needs guidance from religious leaders, governments, and civil society.
